Designing an unsupervised learning architecture to reconstruct brain magnetic resonance images
Brain magnetic resonance imaging (MRI) is a well-known medical imaging modality used by physicians to diagnose neurological disorders. A professional with great knowledge in the field can analyze the obtained 3D volume to detect abnormalities. The main goal of this project is to design and develop a...
| Author: | |
|---|---|
| Format: | master thesis |
| Publication Date: | 2021 |
| Country: | España |
| Institution: | Universitat Oberta de Catalunya (UOC) |
| Repository: | O2, repositorio institucional de la UOC |
| OAI Identifier: | oai:openaccess.uoc.edu:10609/127053 |
| Online Access: | http://hdl.handle.net/10609/127053 |
| Access Level: | Open access |
| Keyword: | autoencoder image reconstruction brain magnetic resonance imaging T1-weighted residual blocks deep learning skipped connections MRI reconstrucció imatges imatges de ressonància magnètica cerebral blocs residuals connexions omeses aprenentatge profund RM reconstrucción de imágenes resonancia magnética cerebral bloques residuales aprendizaje profundo conexiones omitidas |
| Summary: | Brain magnetic resonance imaging (MRI) is a well-known medical imaging modality used by physicians to diagnose neurological disorders. A professional with great knowledge in the field can analyze the obtained 3D volume to detect abnormalities. The main goal of this project is to design and develop a neural network model capable of reconstructing brain MRIs by using an unsupervised approach. In essence, it has to give as result the healthy version of the brain used as input. Therefore, the model has to learn the intrinsic patterns from the healthy brains in order to transfer them to the output. For that reason, T1-weighted images from the IXI Dataset have been used to train our models only with pathology-free images. To ensure the model was learning the desired patterns, not only copying the input directly into the output, the data has been corrupted by applying different types of noise and by masking-out regions. Motivated by the autoencoders simplicity, and with the intention of generating the resulting images with better quality, some experiments have been conducted using different loss functions and trying with different types of skipped connections. The use of both residual blocks and shortcuts to skip long connections has been the combination that has led to the best results. |
|---|